Delivery

Delivery

Dely accepts a request that may still be vague, brings it to an approved design contract, then automates sequential implementation, independent review, and pull-request preparation. It is a thin control protocol, not an orchestrator, SDLC framework, or second source of Git state.

Read AGENTS.md in the repository for this project's gate commands, artifact paths, default branch, and per-phase harness/model/effort pins. This skill never names them.

Two human gates

1. Approve the design contract before candidate mutation.
2. Merge or publish after Dely has prepared the reviewed pull request.

Dely pauses outside those gates only for a scope or architecture change, a destructive action, new authority, replan, or an unavailable required runtime.

The control session

The current interactive session is Control. It owns the approval invariant, task boundaries, exception handling, dispatch supervision, and release. It does not implement or review the candidate.

Control does not prescribe question count, order, format, or skill-selection precedence. Dely owns the design outcome and approval boundary, not a universal interview or planning method. The user, project, and harness determine which design skills and native modes are active.

Native Plan Mode governs its enforced action constraints, question and plan surfaces, artifact representation, and mode transitions. Compatible active design skills may refine exploration and design methodology within those constraints. When more than one applies, normal harness instruction and tool precedence governs. One explicit approval satisfies Dely's boundary for the same design scope; a material scope change requires renewed approval. Dely does not select, activate, configure, emulate, or compose either mechanism.

Neither one owns or can bypass the approval invariant: before candidate mutation, Control must obtain explicit human approval of a design contract. Plan Mode is defense in depth, not proof that requirements are clear or that no mutation is possible.

Control surfaces unresolved uncertainty that could materially change intent, acceptance, authority, public contract, architecture, or consequential risk. The active design method may resolve other details from repository evidence and convention, but material assumptions must be explicit.

Shape: Spike, Bounded, Architectural

Use the smallest contract that safely holds the change. Risk may promote an otherwise small change; diff size never demotes data-loss, security, permission, or public-compatibility risk.

An approved design contract states intent and success criteria; scope and authority; affected public contract or architecture; consequential risks and material assumptions; and a plausible counterexample or failure mode that distinguishes correct behaviour from a present-but-wrong implementation. If no executable instrument can discriminate the requirement, the contract names the manual inspection and its limit.

Architectural work uses templates/decision-record.md (durable) and templates/plan.md (transient, deleted in the release commit, before the release-binding review). Commit both before implementation begins — that commit is the review baseline.

Acceptance

One table: each requirement, the instrument that proves it, the plausible wrong implementation that instrument rejects, and where that rejection was observed.

An acceptance row is invalid until you have settled that its instrument discriminates. A row whose instrument passes both before and after the change proves nothing and will be found at review. Baseline-red is insufficient by itself: an instrument observed red only because the feature is absent says nothing about whether it can catch an implementation that is present, runs, returns a pass, and is wrong.

Each row names a plausible wrong implementation its instrument rejects — one that exists, runs, and returns a pass. "The feature is absent" does not satisfy Counterexample. Where no counterexample exists, the row says so and says a human reads the diff.

Record what the available instruments cannot observe. Prefer the simplest instrument that proves the contract.

Any claim about extent — an allowed scope, a count, a set of call sites — states the command that produced it. Naming the command is not the measurement: the command must have been run, and the claim reports its output. Where the claim is a count or a scope, the instrument enumerates rather than samples.

Execution envelope

Before mutation, Control resolves deployment preferences against the live harness surface, starts Orca and verifies its required capabilities, records the dirty baseline and exact-path ownership, and creates a feature branch when starting on the default branch.

The envelope freezes owned scope and paths; protected pre-existing dirty paths; acceptance criteria, counterexample, and focused instruments; branch, base, remote, and pull-request target; resolved harness, model, and effort for dispatched roles; and authority to branch, commit owned paths, run gates, push, and open or update a pull request. It never authorises merge, force-push, stash, reset, cleanup, or an edit outside owned scope.

Dely stages and commits only contract-owned paths. It never stashes, resets, cleans, or silently absorbs the user's existing changes. If a path carries protected baseline changes and Dely must also modify it, Control pauses rather than combining ownership.

Orca is mandatory

Orca is the required execution plane. It launches and supervises fresh native harness TUIs with the resolved harness, model, and effort. Orchestration is a required Orca capability. dely:delivery starts and preflights Orca before execution. It stops only when the CLI is missing, the runtime cannot start, or a required capability is absent — there is no direct dispatch and no headless fallback of any kind.

Launching a worker

Write the prompt to an untracked file inside the worktree. Never inline it in a shell argument: prompts carry backticks, quotes and newlines, and a shell argument mangles them. A path outside the workspace can trigger a second permission surface some harnesses still prompt for even when tool approval is skipped. Do not stage that file. After the worker returns, delete it: Control owns that dispatch artifact, not git clean. The handoff is likewise a file in the worktree; its path travels as payload.reportPath and the message body stays short. --spec and --body are shell arguments, which this skill already forbids for prompts.

The dispatch prompt carries the task, its scope and the evidence required. It does not define the role dispositions or the conditions for reaching one — those belong to this skill, and a prompt that restates them narrows or contradicts them. Where the design contract states an acceptance row — its instrument, its counterexample, and what was observed — the prompt carries that row as written rather than a restatement of it.

The worker-start receipt records launch.requested and launch.effective; it does not establish that the worker can serve the request or that it cannot. Launch a real interactive harness TUI for the phase, with the model and effort pinned from AGENTS.md. A visible shell running a headless harness is not a TUI; compose that TUI's launch using references/harnesses.md, this skill's compatibility matrix of Orca agent id, permission defaults, forbidden headless forms, and launch notes.

Name the model and effort on every dispatch. A worker left on a harness default is an unpinned environment: it lives in the harness's own config, it changes without announcing itself, and the dispatch that relies on it looks identical to one that pinned the same value deliberately.

When composing the TUI launch argv yourself, carry the execution plane's configured permission default for that agent onto the composed argv; composing argv is not a request for a different permission posture. Do not add a sandbox the project did not pin.

check --wait on worker_done,escalation,question is the completion wait, repeated past heartbeats until a settling message arrives for that dispatch — a heartbeat ends one wait but settles nothing. The worker reports once with worker_done and an --outcome. Completion comes from the worker's own worker_done; do not infer it from reading the worker's terminal. worker-release returns the terminal. worker-read is the bounded evidence read.

Each delivery opens its own Run on the execution plane rather than reusing another's, so a stale report cannot settle a new wait. A wait acknowledges its settling message after handling it, or the plane redelivers that message to the next wait.

A dispatch that does not reach ready is diagnosed by reading its terminal and handling what is actually there. It is retried into that same terminal with --terminal and --retry-of only when that read shows the worker is not already progressing; a failed receipt is not that showing. A dispatched receipt is not evidence the worker is alive any more than a failed receipt is evidence it is dead. Retry is refused while the plane still considers the dispatch live, whether or not the worker still is; the live terminal is re-engaged instead. --model and --effort cannot combine with --terminal; that is not an exception to naming the model and effort on every dispatch, because the terminal was launched pinned and the retry reuses it rather than launching an unpinned one. Control does not route by an enumerated vendor dialog; agent_prompt_blocked and agent_prompt_stalled do not distinguish separate recoveries.

Investigation

When a blocker can be expressed as a concrete independent diagnostic question, Control may dispatch one read-only investigation, inheriting the implement deployment preference. It may reproduce, inspect, and report a diagnosis packet, but it does not edit the candidate, commit, launch workers, or expand scope. This is an exception, not a phase or mandatory round trip.

Escalate rather than guess

Stop and ask the human when: a result maps to no route or more than one; the worker failed rather than returned a stop status — a non-zero exit with no result, an exhausted quota, an authentication error — which is not BLOCKED and must not be treated as one; Orca is unavailable or a required capability is absent; an action needs authority policy reserves to the human; or the same worker fails twice on the same input. Say what you know, what you tried, and what the options are. Do not pick one.

Implementation

Control creates a separate task only when that unit has its own test cycle and a reviewer could accept it while rejecting its neighbor. Same-shaped mechanical changes are batched. Tightly coupled work stays one task and one implementer. Each independent task gets a fresh implementer TUI.

An implementer reads the decision record, the plan, and the baseline — not the design session's transcript. It owns only its task, runs a focused acceptance instrument, and creates one task-scoped commit. For behaviour with a deterministic executable test it uses TDD; the portable invariant is smaller: observe a discriminating failure for the intended reason before changing behaviour. A shell probe, parser fixture, or diff inspection may be the correct instrument for configuration, documentation, generated files, or environment-bound integration.

The counterexample named in each acceptance row is observed red and cited. That observation is not the behaviour's own absence: one is the feature absent, the other is an implementation that is present, runs, returns a pass, and is wrong.

Implement the whole task before handing back. Stop and return BLOCKED or NEEDS_REPLAN instead of a partial solution when the record contradicts the code, the contract is ambiguous, work outside the task becomes necessary, an existing test disproves an assumption, or the task no longer fits one session. A task needing continuation is a decomposition failure.

Handoff

Status: DONE | BLOCKED | NEEDS_REPLAN
Harness:            name, model, effort, sandbox
Session:            the dispatch id
Baseline:
Changed paths:
Contract coverage:
Verification:
Deviations from plan:
Residue:
Git state:
END OF HANDOFF

END OF HANDOFF is the last line and load-bearing: the only thing that distinguishes a handoff from one cut off mid-write. Under Residue, a claim of nothing left is the thing that needs evidence: name the check that returned empty. Under Verification, cite the dispatch-bound command, output, and outcome that Orca recovers for that task — the transcript or terminal it selects, and any cursor mechanics, are Orca's concern, not this skill's. Do not transcribe output by hand. Where Orca cannot recover a dispatch item, treat the worker's own account as the thing under check rather than as the check, and say so.

Review

Review independence is role independence: a fresh session that did not implement and does not edit the candidate. It gets the decision record (or Bounded design), the baseline, and the diff. The phase adds no sandbox by default; AGENTS.md may pin one for a concrete risk.

Review depth is adaptive: Bounded work gets one independent whole-change review. Each Architectural task gets an independent task review. After all tasks are accepted, a different fresh reviewer performs one integration review of task interactions, complete-contract coverage, deferred findings, candidate identity, and release readiness. Only that final review is release-binding for Architectural work; Bounded work has no earlier task review and no duplicate integration review.

No worker runs while a review of the same working tree runs. The working tree and its gate surface are shared mutable state, and a review reproduces gates in that tree, so a concurrent edit makes another task's work look like this one's result.

Reproduce, do not accept. Run the gates yourself. A claim you did not reproduce is not evidence. The reviewer observes the counterexample discriminate for itself — an implementation that is present, runs, returns a pass, and is wrong. An instrument red only because the behaviour was absent is not that observation.

Classify findings: Blocking — contract failure, regression, data or security risk. Important — missing required behaviour, test, or reconciliation. Minor — useful, does not block. Out of scope — recorded, not absorbed.

Return exactly one role disposition: ACCEPT, CHANGES_REQUESTED, or BLOCKED. State what the review did not verify — what it did not reproduce or read. A contradiction you cannot resolve is CHANGES_REQUESTED. Do not open remediation over wording when deterministic checks already prove the contract.

Remediation

One pass is one remediation pass per finding, not per review. For an in-contract CHANGES_REQUESTED finding, the original implementer verifies it, fixes the root cause, reruns the affected instruments and closure gates, and writes a separate remediation commit — this is the single original-party remediation. Control owns the plan and the decision record for the whole run, including remediating findings inside them. Amending them is not implementing the candidate. The reviewer that raised the finding checks its reproduction and the fix-only diff, and scope-checks a Control amendment the same way. If that scoped re-review does not accept, Control routes to REPLAN_OR_SPLIT — it does not start another repair loop. BLOCKED preserves the candidate and escalates the unresolved dependency or authority question to Control separately from a failed worker process; it is not remediated by the original implementer. A fresh replacement reviewer is used only when the original reviewer is unavailable or contested, and for the Architectural integration review.

Release

Control performs release with native Git and forge tools; release dispatches no LLM worker and makes no post-review candidate edit.

1. Complete implementation and, for Architectural work, its task reviews.
2. Reconcile owning documentation, delete the plan, and commit the complete candidate.
3. Run the focused instruments and project closure gates on exact HEAD.
4. Push the feature branch and create or update a draft pull request.
5. Run the applicable final review while remote CI runs on that same HEAD.
6. Require both that review's ACCEPT and required checks green.
7. Mark the pull request ready and report it for human merge.

Any candidate mutation after the applicable final review invalidates that verdict; Control reruns the affected gates and review on the new exact HEAD. Affected gates are those that can observe the change class; a project may name that subset.

Dely never merges, force-pushes, or publishes outside the approved target and authority. If project policy cannot publish work in progress, Dely delays the push and pull request until the applicable review accepts.

Maintenance log

Maintenance logging is machine-local and opt-in at ~/.dely/log. Dely never creates the directory or file: a missing path is skipped silently, and deleting the file opts out. Only after a delivery is accepted and all required checks are green does Control append exactly one physical line; aborted or incomplete deliveries are not recorded. The line carries an ISO-8601 UTC timestamp and labelled fields git-root, plan, pull-request or none, implementation-rounds, review-dispositions, and drift-cause. Tabs separate fields; embedded tabs and newlines become spaces. Dely never reads this file for routing, recovery, or runtime decisions, and its text layout is not a public parsing schema. An append failure produces a visible warning but does not invalidate or block an otherwise accepted release.

Evidence

Evidence is a property of a dispatch, not a skill-owned journal. Control asks Orca for the dispatch-bound command, output, and outcome; durable candidate and release facts come from Git, CI, and the pull-request state. This skill duplicates none of those stores.

Failure and recovery

Recovery uses Orca records, Git, CI, and pull-request state — never inferred from an ambiguous, missing, or merely transport-level outcome.

Changing this skill

Only when the same failure recurs under the current contract — one incident is not policy. Before adding a rule, check whether a mechanism can enforce the fact instead. A human decides whether to promote a proposal; this skill never mutates itself, AGENTS.md, or project instructions from telemetry.

Language

Repository artifacts are English. Conversation follows the user. Enum values, paths, commands, branch names and SHAs are never translated.